Centro de Documentacion de Fundación MAPFRE - Deep learning approach for detecting fake images using texture variation network

Notas: Sumario: Face manipulation technology is rapidly evolving, making it impossible for human eyes to recognize fake faces in photos. Convolutional Neural Network (CNN) discriminators, on the other hand, can quickly achieve high accuracy in distinguishing fake from real face photos. In this paper, we investigate how CNN models distinguish between fake and real faces. According to our findings, face forgery detection heavily relies on the variation in the texture of the images. As a result of the aforementioned discovery, we propose a deep texture variation network, a new model for robust face fraud detection based on convolution and pyramid pooling. Convolution combines pixel intensity and pixel gradient information to create a stationary representation of composition difference information. Simultaneously, multi-scale information fusion based on the pyramid pooling can prevent the texture features from being destroyed. The proposed deep texture variation network outperforms previous techniques on a variety of datasets, including Faceforensics++, DeeperForensics-1.0, CelebDF, and DFDC. The proposed model is less susceptible to image distortion, such as JPEG compression and blur, which is important in this field Registros relacionados: En: Revista Iberoamericana de Inteligencia Artificial. - : IBERAMIA, Sociedad Iberoamericana de Inteligencia Artificial , 2018- = ISSN 1988-3064. - 24/06/2023 Volumen 26 Número 71 - junio 2023 , 14 p.Materia / lugar / evento: Imagen Inteligencia artificial Modelos de análisis Variaciones Otras clasificaciones: 922.134

Derechos: In Copyright (InC)

Referencias externas:

‌ Ver detalle del número